Target detection model construction method based on target data set and secondary transfer learning

Abstract

The application belongs to the technical field of underwater disaster target detection, and particularly relates to a target detection model construction method based on target data set and secondary transfer learning: an acoustic-optical underwater disaster target detection data set is created; a data enhancement method is used to expand the acoustic-optical underwater disaster target detection data set, and meanwhile, a pre-training model weight is used to train a YOLOv5 model on a data set formed by randomly extracting part of the expanded acoustic-optical underwater disaster target detection data set, and the optimal weight in the process is taken as a first-stage weight; the first-stage weight is used to perform two-stage training on the YOLOv5 model on the expanded acoustic-optical underwater disaster target detection data set, and an underwater disaster target detection model is obtained. The model constructed by using the secondary transfer learning and the optical-acoustic data set is used to detect underwater disaster targets, the detection efficiency and accuracy are improved, and the difficulty of detecting underwater disaster targets by using the existing model is reduced.

Description

Technical Field

[0001] This application belongs to the field of underwater disaster target detection technology, and in particular relates to a method for constructing a target detection model based on target dataset and secondary transfer learning. Background Technology

[0002] With the continuous progress and development of human society, the exploration and exploitation of marine resources are becoming increasingly in-depth. Underwater target detection is crucial to the development of marine exploration; therefore, optimizing existing underwater target detection methods is a very important issue. Considering the wide application prospects of deep learning and its potential to further improve the intelligence level of marine exploration, many methods have chosen to introduce deep learning into marine exploration tasks. Currently, many deep learning algorithms serve target detection tasks. Among them, R-CNN, SPP-Net, and Faster RCNN are commonly used two-stage models. In terms of their advantages, they exhibit high accuracy and low false positive rates; however, in terms of their disadvantages, they have high computational complexity, poor real-time performance, and high dependence on annotation data. Currently popular single-stage detection methods include YOLO and SSD. Since they only require one stage to complete the detection task, this determines their high real-time performance and simple, fast characteristics. Compared with two-stage detection algorithms, single-stage algorithms can better achieve a balance between accuracy and real-time performance. Considering the above characteristics, the single-stage detection model YOLOv5 is more suitable for practical application needs. However, when using the single-stage detection model YOLOv5 for target detection, it faces the challenge of obtaining a sufficiently large dataset, thus limiting the detection accuracy and model performance. Furthermore, relying solely on acoustic or optical detection methods cannot simultaneously meet the needs of both near-shore and offshore exploration. Summary of the Invention

[0003] This application provides a method for constructing a target detection model based on a target dataset and secondary transfer learning, which solves the problem of low accuracy in underwater disaster target detection.

[0004] This application provides a method for constructing an object detection model based on a target dataset and secondary transfer learning, including the following steps:

[0005] Step 1: Create an acoustic-optical underwater disaster target detection dataset;

[0006] Step 2: The acoustic-optical underwater disaster target detection dataset is expanded using data augmentation methods. At the same time, a portion of the expanded acoustic-optical underwater disaster target detection dataset is randomly extracted to form a dataset, and the extracted dataset is used to train the YOLOv5 model for one stage. The optimal weights during the training process are used as the weights for the first stage.

[0007] Step 3: Using one-stage weights, the YOLOv5 model is trained in two stages on the expanded acoustic-optical underwater disaster target detection dataset to obtain the underwater disaster target detection model.

[0008] Optionally, in step one, light waves are used to detect shallow water areas and obtain data information of the shallow water area; sonar is used to detect deep water areas and obtain data information of the deep water area. The data information of the shallow water area and the data information of the deep water area constitute an acoustic-optical underwater disaster target detection dataset.

[0009] Optionally, the data augmentation method in step two involves modifying the centralized images in the acoustic-optical underwater disaster target detection dataset.

[0010] Optionally, the methods for modifying the image include rotation, flipping, cropping, resizing, panning, brightness adjustment, contrast adjustment, noise addition, and Gaussian blur.

[0011] Optionally, the YOLOv5 model trained in step two includes the YOLOv5m model.

[0012] Optionally, YOLOv5m uses the CloU loss function to combine the additional losses of the bounding box's scale and aspect ratio. loss The formula is shown below:

[0013]

[0014] In the formula, CIoU loss Let p represent the loss based on the union and intersection of the predicted bounding box and the actual bounding box, p represent the predicted probability of an object appearing in the bounding box, c represent the confidence score of the predicted bounding box, α represent the importance of the confidence score in the loss function, and v represent the variance of the bounding box coordinates.

[0015] Step 3: Using the first-stage weights, the YOLOv5 model is trained in the second stage on the expanded acoustic-optical underwater disaster target detection dataset to obtain the final model weights.

[0016] Optionally, in step three, the YOLOv5m model that has completed the first stage of training is used for the second stage of training on the expanded acoustic-optical underwater disaster target detection dataset. During the training process, the weight parameters of the YOLOv5m model need to be fine-tuned.

[0017] Alternatively, fine-tuning method one: when the acoustic-optical underwater disaster target detection dataset is very similar to the dataset trained in the first stage, the YOLOV5m model uses the features learned in the first stage training to fine-tune all layers.

[0018] Optionally, fine-tuning method two: when the difference between the acoustic-optical underwater disaster target detection dataset and the dataset trained in the first stage is relatively large, freeze some layers of the YOLOV5m model and fine-tune the layers that are not frozen.

[0019] Optionally, the underwater disaster target detection model constructed by this method can be applied to underwater target detection.

[0020] This application provides a method for constructing a target detection model based on a target dataset and secondary transfer learning. First, an acoustic-optical underwater hazard target dataset is created. Traditional data augmentation methods are used to expand the dataset. A one-stage YOLOv5 model is trained on a dataset randomly extracted from the expanded acoustic-optical underwater hazard target detection dataset using pre-trained model weights. The optimal weights obtained during training are used as the first-stage weights. Then, based on the first-stage weights, the model is trained in a second stage on the previously created expanded underwater dataset. This combination of secondary transfer learning and the acoustic-optical dataset improves the model's detection performance and reduces the difficulty of detecting underwater hazard targets using existing models. Compared to using only the acoustic-optical dataset, this method significantly improves detection efficiency and accuracy. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a structural diagram of the YOLOv5 model in this application;

[0023] Figure 2 This is a schematic diagram of the detection results sample of the acoustic-optical underwater disaster target detection dataset of this application;

[0024] Figure 3 A schematic diagram of an optical image sample processed using traditional data augmentation methods;

[0025] Figure 4 A schematic diagram of a sonar image sample processed using traditional data augmentation methods;

[0026] Figure 5-1 This is a confusion matrix generated during the training process using only the acoustic-optical underwater disaster target detection dataset;

[0027] Figure 5-2The graph shows the various loss functions and detection accuracy statistics obtained during the training process using only the acoustic-optical underwater disaster target detection dataset.

[0028] Figure 5-3 Example images of detection results obtained during training using only the acoustic-optical underwater disaster target detection dataset;

[0029] Figure 6-1 The confusion matrix diagram obtained during the training process of the acoustic-optical underwater disaster target detection dataset and the model constructed by secondary transfer learning was used in this application.

[0030] Figure 6-2 This application uses an acoustic-optical underwater disaster target detection dataset and a model built using secondary transfer learning to obtain various loss functions and detection accuracy statistics during the training process;

[0031] Figure 6-3 Example images of detection samples during the training process of the model built using the acoustic-optical underwater disaster target detection dataset and secondary transfer learning in this application;

[0032] Figure 7-1 This is a schematic diagram of the metrics obtained during the validation process using only acoustic-optical datasets.

[0033] Figure 7-2 This is a schematic diagram of the metrics obtained during the verification process of the target detection model constructed using the construction method of this application. Detailed Implementation

[0034] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0035] This application now describes a method for constructing an object detection model based on a target dataset and secondary transfer learning, as provided in its embodiments. The method includes the following steps:

[0036] Step 1: Create an acoustic-optical underwater disaster target detection dataset;

[0037] Step 2: The acoustic-optical underwater disaster target detection dataset is expanded using data augmentation methods. At the same time, a portion of the expanded acoustic-optical underwater disaster target detection dataset is randomly extracted to form a dataset, and the extracted dataset is used to train the YOLOv5 model for one stage. The optimal weights during the training process are used as the weights for the first stage.

[0038] Step 3: Using the first-stage weights obtained in Step 2, perform a second-stage training of the YOLOv5 model on the expanded acoustic-optical underwater disaster target detection dataset to obtain the final model weights, thus obtaining the underwater disaster target detection model.

[0039] Specifically, in step one, light waves are used to detect shallow water areas (coastal regions), forming an optical dataset; sonar is used to detect deep water areas (remote ocean), forming an acoustic dataset; the optical and acoustic datasets together form an acoustic-optical underwater hazard target detection dataset, see [link to relevant documentation]. Figure 2 , which are samples detected using the acoustic-optical underwater disaster target detection dataset.

[0040] Using light waves for shallow water detection solves the problems of complex and unstable sound wave propagation, poor detection accuracy, and low efficiency in shallow water areas. The complex marine environment near the coast, interference from marine life, and frequent human activities severely hinder sound wave propagation, degrading the effectiveness of sonar detection in most coastal waters. Simultaneously, sound wave propagation in shallow water is affected by the shallow water effect, resulting in scattering, refraction, and attenuation, further impacting the effectiveness of acoustic detection in shallow sea areas. Using light waves for shallow water detection effectively solves these problems. It not only clearly depicts the topography and biological conditions in the shallow water environment but also provides real-time image data, facilitating timely and accurate monitoring of the shallow water environment with strong real-time capabilities. Compared to sonar detection, optical detection does not disturb the nearshore marine ecosystem, offering advantages such as non-invasiveness and a green, environmentally friendly approach.

[0041] Sonar is used to explore deep-sea areas (remote ocean). Sound waves not only travel fast in water, helping to cover a wide range of ocean areas and quickly locate and detect deep-sea targets, thus improving the accuracy and efficiency of deep-sea exploration; but also travel long distances in water, reaching tens of thousands of kilometers with minimal attenuation, enabling extensive exploration of deep-sea areas; and compared to other signals, sound waves achieve higher transmission intensity in water, facilitating long-distance detection in the deep sea.

[0042] In step two, the data augmentation method is an existing traditional data augmentation method, which includes performing operations such as rotation, flipping, cropping, resizing, translation, brightness adjustment, contrast adjustment, noise addition, and Gaussian blur on the centralized images of the acoustic-optical underwater disaster target detection dataset.

[0043] Specifically, the optical images in the optical dataset are modified, such as... Figure 3The image shown represents samples processed using traditional data augmentation methods for optical images. By altering the optical images, changes are made to their angle, orientation, size, color, and contrast. Each processed optical image serves as a different input, resulting in a dataset that is several times larger. This image processing enhances the model's generalization ability to the target task by increasing the sample size of the dataset. Simultaneously, the optical dataset suffers from an imbalance in the number of samples across different classes. Therefore, applying the aforementioned data augmentation methods to minority classes increases the number of samples in that class, resolving the imbalance and improving the model's ability to recognize minority class optical images. This enhances the model's robustness, enabling accurate optical image recognition and analysis of the target under varying lighting conditions. Training a YOLOv5 model using this augmented optical dataset significantly improves the model's target detection capabilities.

[0044] The same modification processing described above for optical images in optical datasets also applies to sonar images in acoustic datasets, such as... Figure 4 The image shown is a sample of a sonar image processed using traditional data augmentation methods.

[0045] By combining acoustic and optical detection, the performance of marine environment detection models can be greatly improved, laying a solid foundation for more accurate and efficient detection of the marine environment.

[0046] The first-stage model in step two, YOLOv5, uses the YOLOv5m model it contains. This model is smaller, has a faster detection speed, and can also achieve relatively high detection accuracy.

[0047] See Figure 1 The YOLOv5m model consists of: an input layer (input), a backbone network (backbone), a neck network (neck), and a head network (head).

[0048] YOLOv5m utilizes the CloU (Complete Cross and Union) loss function, incorporating additional losses related to the bounding box's scale and aspect ratio. This enhancement ensures better alignment between the predicted and actual bounding boxes. (CIoU) loss The formula is shown below:

[0049]

[0050] In the formula, CIoU loss denoted as the loss based on the union and intersection (IoU) between the predicted bounding box and the actual bounding box, p represents the predicted probability of an object appearing in the bounding box, c represents the confidence score of the predicted bounding box, α represents the importance of the confidence score in the loss function, and v represents the variance of the bounding box coordinates.

[0051] CIoU loss The function incorporates IoU loss Confidence loss and variance loss are used to optimize the performance of the object detection model, enabling more accurate measurement of the difference between the predicted box and the actual box, and encouraging the model to better predict the object bounding box that is consistent with the actual situation.

[0052] The YOLOv5m model achieves a balance of accuracy, speed, and ease of use. It boasts high-speed real-time performance, maintaining high accuracy while offering faster detection speeds, making it suitable for real-time object detection applications and addressing the shortcomings of previous models in real-time performance. High accuracy is achieved through a deeper model structure and more feature pyramid layers. Multi-scale detection is enabled by utilizing multiple feature pyramid layers and predictions at different scales, allowing for the detection of objects of various sizes, making it suitable for scenarios with objects of different scales. Ease of use is further enhanced by providing user-friendly code and pre-trained weight files, facilitating fine-tuning or customization of object detection tasks.

[0053] The object detection method in this application employs secondary transfer learning to mitigate the impact of the YOLOV5m model's dependence on a large amount of labeled data in the target task. Secondary transfer learning is a deep learning method that transfers knowledge from similar tasks to another, further addressing the problem of excessive reliance on large amounts of labeled data. The first stage of training in secondary transfer learning involves performing a learning task on a task related to the target task (i.e., the acoustic-optical underwater disaster target detection dataset of this application). This learning task is completed by training the model on the target task using a pre-trained weight file. The second stage of training in secondary transfer learning involves training the model on the target task (i.e., the acoustic-optical underwater disaster target detection dataset of this application) based on the optimal weights obtained in the first stage training. The pre-trained model (YOLOV5m model) is an open-source file. Using this file to train the model can significantly reduce training time and difficulty, and improve training accuracy. Utilizing the first-stage weights obtained by adjusting the YOLOV5m model during the previous training process can reduce training time and sample requirements on the target task.

[0054] Through the first stage of training, the YOLOv5m model learns general features and representations useful for the target task. By retraining the already trained YOLOv5m model for the target task, and fine-tuning its parameters during this process, the model's relevance to the target task can be improved. That is, step three, during the second stage of training, allows for fine-tuning of the model according to specific circumstances and requirements. Fine-tuning includes two methods:

[0055] Method 1: When the acoustic-optical underwater disaster target detection dataset is highly similar to the dataset trained in the first stage, the YOLOV5m model is fine-tuned using the features learned in the first stage of training, that is, fine-tuning all layers.

[0056] Fine-tuning all layers includes selecting a base model, freezing weights, adding a new fully connected layer, training the new layer, unfreezing weights, and continuing training. Neural networks consist of multiple layers, each with corresponding weights. This fine-tuning refers to making small adjustments to the weights of each layer when transferring a model to another model if the two models are highly similar.

[0057] Method 2: When there is a relatively large difference between the acoustic-optical underwater disaster target detection dataset and the dataset trained in the first stage, the YOLOV5m model freezes some layers and fine-tunes the unfrozen layers.

[0058] Typically, the model weights of lower-level convolutional layers or feature extraction layers are frozen and not changed; while the model weight parameters of higher-level fully connected layers or classifier layers are fine-tuned.

[0059] The aforementioned fine-tuning process requires data augmentation. In the fine-tuning and training of the acoustic-optical underwater disaster target detection dataset, data augmentation techniques can be applied to increase the diversity of the training data. Random transformations, rotations, and pruning can be applied to the training data to increase the number and diversity of training samples, thereby improving the model's generalization ability.

[0060] The fine-tuned acoustic-optical underwater disaster target detection dataset is validated using a validation set: During fine-tuning or training, a portion of the acoustic-optical underwater disaster target detection dataset is used as a validation set. The validation set can be used to adjust the model's hyperparameters and training strategies. For example, training parameters such as the optimal learning rate, batch size, and number of training iterations can be adjusted during training based on the validation results to achieve optimal training results.

[0061] Because the dataset used in the first-stage training was a subset selected from the expanded underwater disaster target detection dataset, and its similarity to the acoustic-optical underwater disaster target detection dataset used in the second-stage training was very high, a method was adopted to fine-tune the weight parameters of all layers of the YOLOv5m model obtained from the one-stage training. In the actual task execution, the weights from the official pre-trained model file "yolov5m.pt" were used for the first-stage training of the model. The optimal weights from the first-stage training were then transferred to the training on the acoustic-optical underwater disaster target detection dataset, and the weights on all neural network layers were fine-tuned using a method that fine-tuned all layers, thus obtaining a final model with strong recognition capabilities for specific underwater targets.

[0062] In this application, the acoustic-optical underwater hazard target detection dataset was used for all training and testing processes. The difference between the two comparative experiments is that one experiment used the acoustic-optical underwater hazard target detection dataset and a secondary transfer learning method, while the other experiment used only the acoustic-optical underwater hazard target detection dataset for training. The experiments were conducted on a personal computer equipped with an NVIDIA GeForce RTX 4060 and a Max-Q Design 6144MiB graphics card. The detection data obtained during the training process for both experiments are as follows: Figure 5-1 , Figure 5-2 , Figure 5-3 , Figure 6-1 , Figure 6-2 , Figure 6-3 As shown in Figures 7-1 and 7-2, the index graphs obtained during the verification process for the two experiments are shown in Figures 7-1 and 7-2. Figure 5-1 , Figure 5-2 and Figure 5-3 The detection data was obtained by training using only the acoustic-optical underwater hazard target detection dataset. Figure 6-1 , Figure 6-2 and Figure 6-3 This is the detection data obtained from the training process using an acoustic-optical underwater disaster target detection dataset and a secondary transfer learning method; Figure 7-1 and Figure 7-2 The figures are the index curves obtained during the verification process of the two comparative experiments.

[0063] in, Figure 6-2In this context, the following parameters are defined: Confusion matrix, Wrecks, Background, Train / box loss, Train / obj loss, Train / cls_loss, etrics / precision, Metrics, Val / box loss, Val / obj loss, Val / cls_loss, Metrics / mAP_0.5, Metrics / mAP_0.5:0.95, and Metrics / mAP_0.5:0.95.

[0064] Detection samples and data were obtained during the validation process. The data obtained during validation differs somewhat from the data obtained during training. To more intuitively illustrate the differences between the different experimental groups, various metrics obtained during training are listed in Table 1. For convenience, the group using only the acoustic-optical dataset is designated as the control group, while the group using the acoustic-optical underwater disaster target detection dataset and the secondary transfer learning method is designated as the experimental group. The F1 confidence curve, precision-recall curve, precision-confidence curve, and recall-confidence curve obtained during the validation process are shown below. Figure 7-2 As shown.

[0065] Table 1 Data for different groups

[0066] control group 0.779 0.592 0.641 0.412 experimental group 0.793 0.621 0.671 0.439

[0067] Comparing the results obtained from different groups mentioned in Table 1, it can be seen that, in terms of accuracy and efficiency in underwater target detection, the results obtained by training using the secondary transfer learning method and the acoustic-optical dataset are superior to those obtained by training solely on the acoustic-optical underwater disaster target detection dataset, especially in terms of F1 confidence and recall. Test results show that when using the secondary transfer learning method and the acoustic-optical dataset, compared to the method using only the acoustic-optical dataset, precision / confidence can be improved by 1.79%, recall by 4.90%, mAP@0.5 by 4.68%, and mAP@50-95 by 6.55%. The F1 confidence score combines precision and recall into a single metric, providing a balanced measure of the YOLOv5m model's performance. A higher mAP@0.5 indicates that the model can accurately detect and identify objects with higher confidence, meaning that the model has a better ability to distinguish objects from the background. A higher mAP@50-95 indicates that the model performs well in accurately detecting and identifying objects across a wide range of IoU thresholds. The results above demonstrate that the model maintains high accuracy and recall even when there is varying degrees of overlap between the predicted and ground truth bounding boxes. Therefore, the results suggest that the experimental group data exhibits higher localization accuracy and is less affected by environmental uncertainties during detection.

[0068] In the results, since the only target labeled during training was debris, this can also be understood from the classification loss calculation formula, which is as follows:

[0069] C loss =-l*log(p)+(l-1)*log(1-p) (2)

[0070] In (2), l refers to the label of the sample, with a value of 1 for positive samples and 0 for negative samples. p refers to the probability of a positive sample, and since only positive samples exist in each image, the value of p is always 1. Therefore, C... loss Always 0.

[0071] The underwater disaster target detection model constructed using the model building method of this application is applied to underwater target detection, improving detection efficiency and accuracy.

[0072] This application provides a method for constructing a target detection model based on a target dataset and secondary transfer learning. It improves detection performance and reduces the difficulty of detecting underwater disaster targets using existing models by combining secondary transfer learning with an acoustic-optical underwater disaster target detection dataset. First, an acoustic-optical underwater disaster target detection dataset is created. The dataset is expanded using traditional data augmentation methods. A one-stage model, YOLOv5, is trained on a dataset randomly selected from the expanded dataset, and its optimal weights are used as the first-stage weights. Then, the first-stage weights are used to perform a second-stage training on the previously created expanded acoustic-optical underwater disaster target detection dataset to obtain the final model. Test results show that using secondary transfer learning and the acoustic-optical dataset improves the model's detection performance and reduces the difficulty of detecting underwater disaster targets using existing models. Compared with using only the acoustic-optical dataset, it improves detection efficiency and accuracy, increasing precision by 1.79%, recall by 4.90%, mAP@0.5 by 4.68%, and mAP@50-95 by 6.55%.

[0073] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for constructing an object detection model based on a target dataset and secondary transfer learning, characterized in that, Includes the following steps: Step 1: Create an acoustic-optical underwater disaster target detection dataset; Step 2: The acoustic-optical underwater disaster target detection dataset is expanded using data augmentation methods. At the same time, a portion of the expanded acoustic-optical underwater disaster target detection dataset is randomly extracted to form a dataset, and the extracted dataset is used to train the YOLOv5 model for one stage. The optimal weights during the training process are used as the weights for the first stage. Step 3: Using one-stage weights, the YOLOv5 model is trained in two stages on the expanded acoustic-optical underwater disaster target detection dataset to obtain the underwater disaster target detection model. In step one, light waves are used to detect shallow water areas and obtain shallow water data; sonar is used to detect deep water areas and obtain deep water data. The shallow water data and deep water data together constitute an acoustic-optical underwater disaster target detection dataset.

2. The method for constructing a target detection model based on the target dataset and secondary transfer learning as described in claim 1, characterized in that, The data augmentation method in step two involves modifying the images in the acoustic-optical underwater disaster target detection dataset.

3. The method for constructing a target detection model based on the target dataset and secondary transfer learning as described in claim 2, characterized in that, Methods for modifying images include rotation, flipping, cropping, resizing, panning, brightness adjustment, contrast adjustment, noise addition, and Gaussian blur.

4. The method for constructing a target detection model based on the target dataset and secondary transfer learning as described in claim 1, characterized in that, The YOLOv5 model trained in step two includes the YOLOv5m model.

5. The method for constructing an object detection model based on the target dataset and secondary transfer learning as described in claim 4, characterized in that, YOLOv5m uses the CIoU loss function to combine the additional losses of the bounding box's scale and aspect ratio. The formula is shown below: （1） In the formula, This represents the loss based on the union and intersection of the predicted bounding box and the actual bounding box. This represents the predicted probability of an object appearing in the bounding box. This represents the confidence score of the predicted bounding box. This indicates the importance of the confidence score in the loss function. This represents the variance of the bounding box coordinates.

6. The method for constructing a target detection model based on the target dataset and secondary transfer learning as described in claim 1, characterized in that, When the acoustic-optical underwater disaster target detection dataset has a high similarity to the dataset trained in the first stage, the weights of the YOLOV5m model obtained in the first stage of training are fine-tuned across all layers.

7. The method for constructing a target detection model based on the target dataset and secondary transfer learning as described in claim 1, characterized in that, When there is a large difference between the acoustic-optical underwater disaster target detection dataset and the dataset trained in the first stage, the weights of the YOLOV5m model obtained in the first stage training are frozen in some layers, and the unfrozen layers are fine-tuned.

8. The method for constructing a target detection model based on the target dataset and secondary transfer learning as described in claim 1, characterized in that, The underwater disaster target detection model constructed using this method is applied to underwater target detection.

Images